Stripe's $7 Billion Bet on the AI Router

Stripe just agreed to pay more than $7 billion for a company doing roughly $50 million a year in revenue. That is something like a 140 times revenue multiple, four months after the same company was valued at $1.3 billion. Something in that gap is either speculative froth or a genuinely different bet than the profit-and-loss statement suggests. I think it is mostly the second thing, with a real dose of the first.

The company is OpenRouter, an API gateway that lets a developer call one endpoint and get routed to whichever of more than 400 AI models fits the job, OpenAI, Anthropic, Google, xAI, DeepSeek, all behind a single bill. Bloomberg confirmed the finalized deal on August 16 and TechCrunch reported the terms: more than $7 billion, up from the $1.3 billion valuation OpenRouter set in May with a $113 million Series B led by Alphabet's CapitalG, as TechCrunch covered at the time. The Wall Street Journal had reported talks at a figure closer to $10 billion before negotiations settled lower, which tells you this moved fast and the final number still landed far above where the company traded three months earlier.

I have spent most of this week writing about AI prices falling. This deal is the flip side of that story: a bet that the company sitting between buyers and that price war is worth more than the price war itself. That is not a crazy idea. It is the same logic that makes an exchange worth more than most of what trades on it. If model prices keep swinging, DeepSeek cutting one week and a frontier lab slashing a tier the next, then whoever lets a business shop between models without rewriting its code is doing something genuinely useful, and useful things that touch a lot of traffic get expensive fast.

Here is the strongest case for paying up. OpenRouter is not small. It was moving something like 100 trillion tokens a month by mid-2026, five times its volume from six months earlier, across 8 million users. Stripe was arguably the best-positioned buyer that exists: OpenRouter already processes its credit purchases through Stripe, per OpenRouter's own pricing documentation, so Stripe had a paid-up view of the growth curve before it made an offer. And Forkast's analysis of the deal notes that rivals including Databricks, Cursor, and Ramp are reportedly building competing routing features. Buying the category leader, with real usage data on which model wins which task, before a competitor commoditizes the category, is a defensible reason to move fast and pay a premium. Waiting has a cost when a market this new is this contested.

Now look at what that fee actually is. OpenRouter passes through provider pricing with zero markup on the model itself, by its own account. It earns roughly 5 percent on credit purchases, and, past a $25,000 monthly allowance, 5 percent on usage where customers bring their own provider keys. Research firm Sacra put its annualized revenue at about $50 million as of early 2026, up from $19 million a year earlier, genuinely strong growth of around 163 percent according to Sacra's OpenRouter research. Strong growth on a small revenue base is still a small revenue base, and that 5 percent take is tied directly to the trend I keep documenting on this blog: falling per-token prices. Every price cut from a major lab shrinks the dollar amount that 5 percent applies to, even as usage climbs. OpenRouter's volume has to keep outrunning the price declines just to hold revenue flat, and 2026 has been one long proof that those declines are not stopping.

Here is the number that matters more to a business owner than to Stripe's board. If your company spends $50,000 a month on inference through a bring-your-own-key gateway like this, you are paying 5 percent, about $1,250 a month, on the $25,000 above the free allowance. That is real money for what is, underneath the branding, a routing decision: which model handles which request, and when to fail over to a cheaper one. A team with meaningful volume and one engineer who understands the workload can build that logic in a matter of days, not months, and keep the fee for itself. Use a router-as-a-service while your spend is small and speed matters more than the fee. Once the inference bill has real weight, the arithmetic favors owning the routing.

That is the build-versus-rent call we walk clients through constantly, and this deal is a useful data point either way: the market just valued the routing layer at billions in aggregate, which is exactly why it is worth owning yourself once your volume clears the threshold where the engineering hour pays for itself. If you want a second opinion on whether your AI spend has crossed that line, reach out.